A brief introduction to Pastry

1
A brief introduction to Pastry

• Based on
• A. Rowstron and P. Druschel, Pastry Scalable,
  decentralized object location and routing for
  large-scale peer-to-peer systems, Middleware,
  Nov. 2001.
• Presenter Fabián E. Bustamante

2
Distributed Systems Architectures

• System architecture Component placement and
  their interactions
• Client/Server
• Server implementing a specific service
• Client requests the service
• Peer-to-Peer
• All peers are equal
• Symmetric interaction between processes
• Focusing on data-sharing systems

3
Data Sharing P2P

• Data sharing P2P apps rely on overlay network
  protocols for object storage/retrieval message
  routing
• Classes of overlay protocols
• Unstructured
• Few constraints on overlay construction data
  placement
• Could support arbitrary complex queries, highly
  resilient to churn
• Restricted to inefficient, near-blind search
  strategies
• Structured (DHT Distributed Hash Tables)
• Constraining overlay structure data placement
• Efficient object discovery
• Potential problems handling churn, exploiting
  node heterogeneity supporting complex queries

4
Pastry DHT

• Pastry
• Scalable, decentralized object location and
  routing for large-scale P2P systems
• Nodes assigned a unique 128-bit numeric node
  identifier (nodeId) (circular id space)
• NodeID can be expressed in based 2b, b 4
• Given a msg its 128-bit key, Pastry routes the
  msg to the node with numerically closest nodeId
  (in less than )
• At each step, forward to node sharing with a key
  a 1 larger prefix or is numerically closer

5
Pastry Node State

• Each Pastry node maintains
• Routing table
• Log N rows with 2b - 1 entries
• Entries at row n share n digit with current node
• Choice of b involves a trade-off between size of
  populated portions of the routing table maximum
  number of hos to route between two nodes (e.g. b
  4 106, avg. 75 entries and 5 hops)
• Neighborhood set
• nodeIds and IP addresses of M nodes closest
  (proximity metric) to the local node (M 2
  2b)
• Leaf set
• Set of L/2 numerically closest larger/smaller
  nodeIds relative to the present nodes nodeId
  (L 2b)

6
Routing in Pastry

• Routing a message with key D arriving at a node
  with nodeId A
• If (D is within range of leafset)
• Forward to Li such that D - Li is minimal
• Else // use routing table
• Let l be the length of prefix shared bet/ A and D
• If ! null
• Forward to
• Else
• Forward to T in L U R U M such that the length of
  prefix between T and D is gt l and T - D lt A -
  D

7
Pastry API

• Pastry exports (downcalls)
• nodeId pastryInit(Credentials, Applications)
• To join the local node (application is a handle
  to a set of procedures to invoke when certain
  events happen)
• route(msg, key)
• Applications must export (upcalls)
• deliver(msg, key)
• Call by Pastry when msg arrived for node
• forward(msg, key, nextId)
• Call by Pastry just before forwarding the msg to
  node with nodeId nextId
• newLeafs(leafSet)
• Call by Pastry whenever theres a change in local
  node leafset

8
Dynamics Node Arrival

• Node X knows of nearby node A
• X asks A to route a join msg with key X, A routes
  it to node Z
• In response to join msg, A, Z and all in between
  send Z their state tables
• X initializes its own state tables
• As neighborhood set is Xs
• Zs leafset is basis for Xs leafset (closest
  nodeId)
• Xs R0 (X0) A0, X1 B1, where B is the first
  node in the route from A to Z
• X informs everyone who should know of its arrival

9
Example Applications

• PAST A persistent storage service
• fileId a function of files name owner
• Replicas of a file are stored on the k Pastry
  nodes with nodeId numerically closest to fileId
• Scribe A publish/subscribe system
• Subscribers are kept in node with nodeId
  numerically closest to topicId)
• Others
• SplitStream
• Squirrel